sched: fix inconsistency when redistribute per-cpu tg->cfs_rq shares
authorKen Chen <kenchen@google.com>
Wed, 19 Nov 2008 06:41:57 +0000 (22:41 -0800)
committerIngo Molnar <mingo@elte.hu>
Wed, 19 Nov 2008 17:39:37 +0000 (18:39 +0100)
Impact: make load-balancing more consistent

In the update_shares() path leading to tg_shares_up(), the calculation of
per-cpu cfs_rq shares is rather erratic even under moderate task wake up
rate.  The problem is that the per-cpu tg->cfs_rq load weight used in the
sd_rq_weight aggregation and actual redistribution of the cfs_rq->shares
are collected at different time.  Under moderate system load, we've seen
quite a bit of variation on the cfs_rq->shares and ultimately wildly
affects sched_entity's load weight.

This patch caches the result of initial per-cpu load weight when doing the
sum calculation, and then pass it down to update_group_shares_cpu() for
redistributing per-cpu cfs_rq shares.  This allows consistent total cfs_rq
shares across all CPUs. It also simplifies the rounding and zero load
weight check.

Signed-off-by: Ken Chen <kenchen@google.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched.c

index a4c156d9a4a56d65d0e081a95dc213ffd76060ec..93bfb086e60fe66f37bdf1a4b87c040b7c740be1 100644 (file)
@@ -1453,27 +1453,13 @@ static void
 update_group_shares_cpu(struct task_group *tg, int cpu,
                        unsigned long sd_shares, unsigned long sd_rq_weight)
 {
-       int boost = 0;
        unsigned long shares;
        unsigned long rq_weight;
 
        if (!tg->se[cpu])
                return;
 
-       rq_weight = tg->cfs_rq[cpu]->load.weight;
-
-       /*
-        * If there are currently no tasks on the cpu pretend there is one of
-        * average load so that when a new task gets to run here it will not
-        * get delayed by group starvation.
-        */
-       if (!rq_weight) {
-               boost = 1;
-               rq_weight = NICE_0_LOAD;
-       }
-
-       if (unlikely(rq_weight > sd_rq_weight))
-               rq_weight = sd_rq_weight;
+       rq_weight = tg->cfs_rq[cpu]->rq_weight;
 
        /*
         *           \Sum shares * rq_weight
@@ -1481,7 +1467,7 @@ update_group_shares_cpu(struct task_group *tg, int cpu,
         *               \Sum rq_weight
         *
         */
-       shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
+       shares = (sd_shares * rq_weight) / sd_rq_weight;
        shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
 
        if (abs(shares - tg->se[cpu]->load.weight) >
@@ -1490,11 +1476,7 @@ update_group_shares_cpu(struct task_group *tg, int cpu,
                unsigned long flags;
 
                spin_lock_irqsave(&rq->lock, flags);
-               /*
-                * record the actual number of shares, not the boosted amount.
-                */
-               tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
-               tg->cfs_rq[cpu]->rq_weight = rq_weight;
+               tg->cfs_rq[cpu]->shares = shares;
 
                __set_se_shares(tg->se[cpu], shares);
                spin_unlock_irqrestore(&rq->lock, flags);
@@ -1508,13 +1490,23 @@ update_group_shares_cpu(struct task_group *tg, int cpu,
  */
 static int tg_shares_up(struct task_group *tg, void *data)
 {
-       unsigned long rq_weight = 0;
+       unsigned long weight, rq_weight = 0;
        unsigned long shares = 0;
        struct sched_domain *sd = data;
        int i;
 
        for_each_cpu_mask(i, sd->span) {
-               rq_weight += tg->cfs_rq[i]->load.weight;
+               /*
+                * If there are currently no tasks on the cpu pretend there
+                * is one of average load so that when a new task gets to
+                * run here it will not get delayed by group starvation.
+                */
+               weight = tg->cfs_rq[i]->load.weight;
+               if (!weight)
+                       weight = NICE_0_LOAD;
+
+               tg->cfs_rq[i]->rq_weight = weight;
+               rq_weight += weight;
                shares += tg->cfs_rq[i]->shares;
        }
 
@@ -1524,9 +1516,6 @@ static int tg_shares_up(struct task_group *tg, void *data)
        if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
                shares = tg->shares;
 
-       if (!rq_weight)
-               rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
-
        for_each_cpu_mask(i, sd->span)
                update_group_shares_cpu(tg, i, shares, rq_weight);